Consistent Response Generation with Controlled Specificity

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4418–4427 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 4418 Consistent Response Generation with Controlled Specificity Junya Takayama and Yuki Arase Graduate School of Information Science and Technology, Osaka University {takayama.junya, arase}@ist.osaka-u.ac.jp Abstract We propose a method to control the speci- ficity of responses while maintaining the con- sistency with the utterances for open-domain conversation systems. We first design a met- ric based on pointwise mutual information, which measures the co-occurrence degree be- tween an utterance and a response. To con- trol the specificity of the generated responses, we add the distant supervision based on the co- occurrence degree and a PMI-based word pre- diction mechanism to a sequence-to-sequence model. Using these mechanisms, our model outputs the words with desired specificity for a given specificity level. In experiments with open-domain dialogue corpora, automatic and human evaluation results confirm that our model controls the specificity of the responses more sensitively than the conventional model and can generate highly consistent responses. 1 Introduction Open-domain response generation is a task for gen- erating a human-like responses to chit-chatting. There are many end-to-end response generation models (Vinyals and Le, 2015; Sordoni et al., 2015; Mei et al., 2017) that apply a sequence-to-sequence (Seq2Seq) (Sutskever et al., 2014) architecture, which allows the generation of fluent responses. However, the Seq2Seq model suffers from a ten- dency to generate safe but overly typical responses (i.e. dull responses), such as “Yes” and “I don’t understand.” To solve this problem, several studies proposed methods to increase the specificity of the generated responses (Li et al., 2016a; Zhang et al., 2018b; Jiang et al., 2019); however, simply maxi- mizing the specificity of the response results in a degenerative solution that generates a specific but inconsistent responses. In this study, we define the conditions that an automatically generated response is expected to sat- isfy as (i) being consistent with an input utterance, It's cold outside… Yes Yeah, it's cold Be careful not to catch a cold High Low Specificity Utterance Possible responses Figure 1: An example of the relationship between ut- terance and response. There are several possible re- sponses to an utterance with various specificity. (ii) being specific to provide informative contents, and (iii) being controllable. As shown in Figure 1, in a human conversation, an utterance could have various responses with different specificity (Cs´aky et al., 2019). Then, humans control the specificity of the response as necessary. Thus, instead of only generating highly specific responses, the specificity should be controllable in response generation tasks. We propose a method to control the specificity of responses while maintaining their consistency with the utterances. Following the observation that a re- sponse uniquely co-occurring with a specific utter- ance in a corpus is both specific and consistent for the utterance, we design a metric called MaxPMI, which measures the co-occurrence degree between an utterance and a response on the basis of positive pointwise mutual information (PPMI). We apply the distant supervision …

Content truncated to 500 words